Pinned
For the record: attention, scale, and a sufficiently hard problem is all you need
Kyle Vedder
1,390 posts
Kyle Vedder
@KyleVedder
- They actually have 21,000 students, including grad students. Most admin are for non-undergrad related activities, e.g. contract fulfillment as an R1 institution. Here's a breakdown by function. I agree there's bloat, but # undergrad is not the full story. oir.harvard.edu/fact-book/facu…
- One of Pi0's novel architecture bits is the use of a Flow Matching action head -- previous to this, modern VLAs like OpenVLA leveraged diffusion diffusion heads What is a flow matching head? What makes it easier to use versus other denoising heads? A short thread!🧵 (1/7)
GIF - Neural Eulerian Scene Flow Fields EulerFlow is - unsupervised - state of the art for point cloud scene flow - flexible (works without tuning across domains) - canonical frame-free - also a point tracker a thread 🧵👇
GIF
GIF - As promised: three commercially valuable tasks continuously demoed live for all attendees Public demos are a big deal -- there's no reshoots or camera tricks to hide flaws Our stack is simply robust enough to set up multiple bots and showcase valuable manipulation skills live
00:00
>nonstop live dexterous manipulation demo we're gonna blow these ppls minds this is the kind of thing you can't even see at robotics conferences (yet, stay tuned 😉) - just finished the last day of my internship at Nvidia 🫡 stay tuned, will be dropping some honestly incredible results from my project soon
- Carmack fell for level 1 ML trope 😔 biological brains have had many MORE orders of magnitude of computer poured into them via joint neural architecture search and weight optimization overparameterization of ML models enables this search (with gradients!) for lottery ticketsYes, just not in LLMs. We have a clear biological existence proof that the training scale of frontier LLMs is unnecessary.
- Replying to @chris_j_paxtonas a matter of transparency: this is a demo video aimed at business customers. it is intended to allow them to visualize what a robot could do for their business the various tasks shown in that video were recorded via human teleop. our research videos are autonomous, as labeled
- Replying to @dylan_ebert_It's not a NeRF (Neural Radiance Field), there's not a neural network. Which is the whole point; using Gaussian Splatting is far faster and higher quality than using a coordinate network.
- Nvidia should have called it Thor's Hammer this is basically a standalone full GPU
- Happy Independence Day 🦅🌎🇺🇸 here's another shot of our entirely autonomous interactive policy despite knowing DYNA-1 can do robust long horizon manipulation, the flag hand-off was a wow moment for me bc it felt like a pet dog (hence the audio 🐶), not alien intelligence
00:00
00:13Happy July 4th from Dyna! 🇺🇸🇺🇸 - I found a surprising new way to use my thesis to support my current work
- - new demo task - new environment with hardware we just setup - robust to changes in lighting, people visible, etc - production ready quality and robustness - real businesses that are interested bc it provides real value it's a cook or get cooked world and we're cookingWe have started taking DYNA-1, our dexterous robust VLA model, to conferences and showcasing it for hours on end! The model run for 3 days, 8 hours each day at #HITEC2025 3 weeks ago with 99.9% overall success rate (dropped 1 towel in day 2). No intervention, it just works :)
00:00
